November brought an uncomfortable reminder. On the 12th, Microsoft 365 suffered an outage affecting Outlook, Teams, SharePoint, OneDrive and Copilot that, according to independent incident trackers, lasted much of the day. On the 15th, AWS’s us-east-1 region had several hours of problems, including sign-in failures that kept people from even reaching their consoles to diagnose.

Neither provider had published a root-cause analysis when I wrote this. But you do not need the exact cause to draw architecture conclusions.

Concentration is the risk

us-east-1 is not just another region: it is AWS’s oldest and most used, and many global services depend on it. Every time it has problems, the impact reaches companies that thought they were “in the cloud” without realizing that, in practice, they were in a single place.

Microsoft 365 shows the other side: when email, collaboration and the AI assistant all depend on the same provider, an outage does not stop a tool, it stops the way people work.

What is new: processes now depend on AI

Until recently, a productivity outage was an inconvenience. Today many organizations are putting AI agents in the middle of business processes: customer service, risk analysis, operations. If the model, the agent platform or the region they run in is unavailable, that process stops.

And the dependency chain of an AI solution is long: the model provider, the orchestration platform, the vector database, internal APIs, the identity provider. Any of those links can be a single point of failure.

Five design decisions

Resilience is not bought; it is designed. These are the decisions I recommend reviewing:

1. Map your real dependencies. Not the ones in the diagram, the real ones. Which region does your identity provider depend on? Where does the model your agent uses run? Many surprises only show up in this exercise.

2. Define what must keep working. Not everything needs high availability. Identify the critical processes and set, for each one, how much downtime is acceptable.

3. Design for degradation, not just recovery. If the AI agent does not respond, can the process continue manually or with simple rules? A system that degrades gracefully is worth more than one that promises never to fail.

4. Do not depend on a single model. An abstraction layer that lets you switch models or providers reduces both operational risk and commercial lock-in. What holds for your database holds for AI.

5. Test it. A contingency plan that has never been run is not a plan. Simulating the loss of a region or an AI provider, in a controlled way, is the only way to know whether the architecture holds.

My take

Outages will keep happening; they are part of running complex systems at global scale. The difference is whether a provider outage becomes a business outage. In the age of AI agents, that question moves from the infrastructure team to the boardroom.


I help organizations assess their architecture and provider dependency before an outage does it for them. See services.

Sources: Control Plane cloud outage tracker; Business Standard on 2026 outages and cyber incidents.